HW-Accelerated Physical AI Framework For Resource-Constrained Edge Devices (ASU)

A new technical paper titled “Enabling Physical AI at the Edge: Hardware-Accelerated Recovery of System Dynamics” was published by researchers at Arizona State University.

Abstract
“Physical AI at the edge—enabling autonomous systems to understand and predict real-world dynamics in realtime—demands efficient hardware acceleration. Model recovery (MR), which extracts governing equations from sensor data, is critical for safe and explainable monitoring in mission-critical autonomous systems (MCAS) operating under severe time, compute, and power constraints. While Field Programmable Gate Arrays (FPGAs) offer promising reconfigurable hardware for edge deployment, state-of-the-art (SOTA) MR methods like EMILY and PINN+SR rely on Neural ODEs requiring iterative solvers that resist hardware acceleration. This paper presents MERINDA (Model Recovery in Dynamic Architecture), an FPGA-accelerated framework specifically designed to enable physical AI at the edge. MERINDA replaces computationally expensive Neural Ordinary Differential Equation (ODE) components with a hardware-friendly architecture combining: (a) Gated Recurrent Unit (GRU) layers for discretized dynamics, (b) dense inverse ODE layers, (c) sparsity-driven dropout, and (d) lightweight ODE solvers—with critical components fully parallelized on FPGA. Evaluated on four benchmark nonlinear dynamical systems, MERINDA achieves transformative improvements over Graphics Processing Unit (GPU) implementations: 114× reduction in energy consumption (434J vs. 49,375J), 28× smaller memory footprint (214MB vs. 6,118MB), and 1.68× faster training—while maintaining model recovery accuracy equivalent to SOTA methods. These results validate MERINDA’s capability to bring physical AI to resource-constrained edge devices for real-time autonomous system monitoring.”

Find the technical paper here. Published December 2025.

Xu, Bin, Ayan Banerjee, and Sandeep Gupta. “Enabling Physical AI at the Edge: Hardware-Accelerated Recovery of System Dynamics.” arXiv preprint arXiv:2512.23767 (2025).